Neural network verification based on input splitting and forward propagation of symbolic intervals with fresh variables.

Note: DPNeurifyFV is currently only supported for x86 architectures.

Open the Julia REPL in this directory and type

• ] to open the Julia package manager
• add https://github.com/sisl/NeuralVerification.jl#0d9be34 (adds NeuralVerification.jl to the environment, but pinned to a specific commit)
  • at least use the pinned version, until this issue with the installation of NeuralVerification.jl is resolved
• add https://github.com/sisl/NeuralPriorityOptimizer.jl and follow the installation instructions on their repo
  • install Gurobi (academic licenses are free)
  • type install Gurobi
  • (to use all features of NeuralPriorityOptimizer, you also need to install Mosek, however, this is not necessary for DPNeurifyFV or the experimental evaluation.)
• add LazySets
• add https://github.com/phK3/DPNeurifyFV.jl

In order to maximize the first output of the ACAS_1_1 network over the input region defined by the first property of the ACAS-Xu benchmark suite run

using NeuralVerification, DPNeurifyFV

input_set, output_set = DPNeurifyFV.get_acas_sets(1)

acas = read_nnet("./networks/ACASXU_experimental_v2a_1_1.nnet")

params = DPNeurifyFV.PriorityOptimizerParameters(max_steps=5000, print_frequency=100, stop_frequency=1, verbosity=2)
optimize_linear_deep_poly(acas, input_set, [1.,0,0,0,0], params, solver=DPNFV(method=:DeepPolyRelax, max_vars=15), concrete_sample=:BoundsMaximizer, split=DPNeurifyFV.split_important_interval)

Further examples can be found in ./example_notebook.ipynb.

To reproduce the experiments in the paper, open the Julia REPL in this directory and type

• ] to open the package manager
• activate . (to activate the DPNeurifyFV environment)
• press backspace to leave the package manager
• type include("experiments/run_experiments.jl") to start verification of the ACAS Xu benchmark set using DPNeurifyFV
  • timeouts etc. can be changed in the script
• type include("experiments/zope_experiments.jl") to start verification with ZoPE
• to reproduce the results for NNENUM,
  • download NNENUM from GitHub: https://github.com/stanleybak/nnenum
  • use the Python script under experiments/run_nnenum.py and place it in nnenum/examples/acasxu/
  • run the script